J. M. VALLS, I. M. GALVÁN, and P. ISASI
Universidad Carlos III de Madrid, Spain
Lazy learning methods [1–3] are conceptually straightforward approaches to approximating real- or discrete-valued target functions. These learning algorithms defer the decision of how to generalize beyond the training data until a new query is encountered. When the query instance is received, a set of similar related patterns is retrieved from the available training patterns set and is used to approximate the new instance. Similar patterns are chosen by means of a distance metric in such a way that nearby points have higher relevance.
Lazy methods generally work by selecting the k nearest input patterns from the query points. Usually, the metric used is the Euclidean distance. Afterward, a local approximation using the samples selected is carried out with the purpose of generalizing the new instance. The most basic form is the k-nearest neighbor method . In this case, the approximation of the new sample is just the most common output value among the k examples selected. A refinement of this method, called weighted k-nearest neighbor , can also be used, which consists of weighting the contribution of each of the k neighbors according to the distance to the new query, giving greater weight to closer neighbors. Another strategy to determine the ...